Understanding matching extraction method and device based on industry large model
Through the understanding matching extraction method based on industry big models, the problems of low efficiency and insufficient accuracy of manual interpretation in traditional methods are solved, and automated semantic understanding and matching extraction are realized, which improves work efficiency and accuracy, and ensures the matching results. Objectivity and accuracy.
Patent Information
- Application Number
- CN202411049452.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2024-11-29
AI Technical Summary
Traditional methods rely on manual interpretation and matching in industry management and decision-making, which has subjective deviations and inefficiency problems. It requires workers to label labels for a lot of time, affecting actual execution efficiency, and cannot achieve the optimal effect and highest in some industries. Accuracy.
Using the understanding matching extraction method based on industry big models, learn the semantic characteristics of special indicators in the industry through smaller models in the big models, set the similarity threshold, and realize the semantic understanding and contextual association of the overall statement, interactive confirmation, and optimal matching. The item is used as the extraction result and triggers the exception processing flow when the data is incomplete.
The semantic understanding and matching extraction of special indicators has been realized, which reduces manual intervention, improves work efficiency, reduces matching error rate, and improves accuracy. The accuracy of experimental tests can reach 98.6%, ensuring that the matching results meet the intentions of industry personnel, reduces subjective deviations, and improves Objectivity of decision-making.
Smart Images

Figure CN119047478A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence technology, and specifically provides a method and device for understanding, matching and extracting based on an industry big model. Background Art
[0002] In industry management and decision-making, accurately understanding industry-specific indicators and matching and extracting them is crucial to improving work efficiency and decision-making accuracy. Traditional methods often rely on manual interpretation and matching, which has subjective bias and low efficiency.
[0003] For existing technical aspects, reference may be made to patent CN112487182B "Training method, text processing method and device for text processing model". This patent provides a training method, text method and device for text processing models in the field of natural language processing in the field of artificial intelligence. The training method comprises: obtaining training text; inputting the training text into a teacher model and a student model respectively to obtain sample data output by the teacher model and prediction data output by the student model, wherein the teacher model and the student model respectively comprise an input layer, one or more intermediate layers and an output layer, the sample data comprises sample semantic features output by the intermediate layer of the teacher model and sample labels output by the output layer, the prediction data comprises prediction semantic features output by the intermediate layer of the student model and prediction labels output by the output layer; the model parameters of the student model are trained based on the sample data and the prediction data to obtain a target student model, thereby improving the accuracy of the text processing results of the student model.
[0004] The difficulty of this technology is that it requires workers to label, which is time-consuming and affects the actual execution efficiency. In addition, the accuracy can only be adjusted by parameters, so it is impossible to achieve the best effect and highest accuracy in some industries.
[0005] In order to solve these problems, how to quickly and accurately understand and extract the semantic information of special indicators in the industry and match it with the big model to provide reliable data support and reference basis for decision makers is a technical problem that technical personnel in this field need to solve urgently. Summary of the invention
[0006] The present invention aims at addressing the deficiencies of the above-mentioned prior art and provides a practical understanding matching extraction method based on a large industry model.
[0007] A further technical task of the present invention is to provide a reasonably designed, safe and applicable understanding, matching and extraction device based on an industry big model.
[0008] The technical solution adopted by the present invention to solve its technical problem is:
[0009] A method for understanding, matching and extracting based on an industry big model has the following steps:
[0010] S1, the smaller models within the large model learn the semantic features of industry-specific indicators;
[0011] S2, set a similarity threshold and extract the top items whose semantic similarity is higher than the threshold;
[0012] S3. The large model realizes the semantic understanding of the whole sentence, including context association and semantic inference;
[0013] S4. According to the semantic understanding result, the best matching item is selected as the extraction result;
[0014] S5. Interactive confirmation to ensure that the matching results meet the intentions of industry personnel;
[0015] S6. If the matching result does not meet the intention of the industry personnel, repeat the above steps to perform matching extraction.
[0016] Furthermore, in step S1, it further includes:
[0017] S1-1. Determine the learning objectives, define the specific problems to be solved, the expected business goals, and the role of the big model in specific application scenarios;
[0018] S1-2. Collect special indicators. Collect relevant special indicators according to learning objectives and large models.
[0019] S1-3. Determine whether the data is complete and conduct a comprehensive check on the collected data, including verifying whether the number of data records meets expectations, whether there are missing values for each special indicator, and whether the data distribution is reasonable.
[0020] S1-4, data preprocessing, if the judgment result shows that the data is complete, then enter the data preprocessing stage;
[0021] S1-5, use a smaller model within the larger model to learn semantic features. After completing data preprocessing, use the smaller model for training to learn the semantic features contained in the special indicators.
[0022] S1-6. Evaluate the effect of the large model. After the large model training is completed, use special indicators to conduct a detailed evaluation of the performance of the large model;
[0023] S1-7, exception handling. If the result of step S3 is that the data is incomplete, the exception handling process is triggered. At this time, the data source should be traced back to find the cause, or the missing data should be supplemented with the help of data interpolation technology. After that, steps S3 to S6 are re-executed to form an iterative loop until the data meets the integrity requirements and the large model learning and evaluation process can proceed smoothly.
[0024] Furthermore, in step S2, first, a similarity threshold is set as a basis for judgment, and then, the semantic similarity between the texts to be compared is calculated using an algorithm;
[0025] Then, a judgment is made. If the calculated similarity is higher than the preset threshold, it indicates that the semantics of the two texts are highly consistent. At this time, relevant indicators can be further extracted;
[0026] On the contrary, if the similarity is lower than the threshold, the two texts are judged to be semantically inconsistent, and the large model feedbacks the conclusion of semantic inconsistency and throws out in detail the abnormal reasons that lead to low similarity.
[0027] Furthermore, in step S3, first, the big model receives a complete sentence provided by an industry practitioner as initial input. Next, the big model performs preliminary semantic and contextual relevance judgments, and evaluates the relevance of the sentence to the known context by considering the sentence's grammatical parsing, word meaning analysis, and potential background knowledge.
[0028] Furthermore, if the judgment result is relevant, the big model will enter the semantic inference stage, using natural language understanding and reasoning technology. If the big model's judgment result shows that the sentence is irrelevant to the current context, the big model does not immediately give up understanding, but instead starts a context search mechanism to find context information related to the input sentence topic from a wider range of Internet industry resources;
[0029] After finding the relevant context, the large model will be integrated with the original context to expand the knowledge background, and then return to step S2 to re-judge the relevance of the semantics with the updated context, forming an iterative loop.
[0030] Furthermore, in step S4, first, the big model obtains the semantic understanding results generated by the previous link, and then, the big model uses a predefined matching algorithm or rule to calculate the matching degree between the understanding results and the internal knowledge base, database or specific target data set;
[0031] If the matching degree does not reach the threshold, it means that there is a large discrepancy between the semantic understanding result and the expected data source. At this time, the large model does not generate extraction results and provides feedback prompts to industry personnel.
[0032] Further, in step S5, first, the big model receives preliminary matching results from the matching module, and the big model presents these matching results to the industry personnel. The big model initiates a confirmation request, asking the industry personnel whether they think these matching results are consistent with the requirements, and the industry personnel give feedback by clicking a yes or no button;
[0033] If the industry personnel confirms that the matching result is accurate, that is, they select Yes, the big model will record the matching process and generate or update the final shopping index of the industry personnel based on the matching result;
[0034] If the industry personnel feedback that the matching result is inaccurate, that is, select "No", the large model will start the loop processing mechanism and re-select other options from the alternative matching results that previously exceeded the preset threshold and show them to the industry personnel again.
[0035] Furthermore, in step S5, regarding exception handling, if a matching result that satisfies the industry personnel is not found during the loop process, or the industry personnel cancel the confirmation multiple times in a row, the large model triggers an alarm mechanism, prompting the operation personnel to intervene and analyze the cause.
[0036] An understanding matching extraction device based on an industry big model comprises: at least one memory and at least one processor;
[0037] The at least one memory is used to store a machine-readable program;
[0038] The at least one processor is used to call the machine-readable program to execute a large model-based understanding matching extraction method.
[0039] Compared with the prior art, the understanding matching extraction method and device based on the industry big model of the present invention has the following outstanding beneficial effects:
[0040] The present invention automates the semantic understanding and matching extraction of special indicators, reduces manual intervention, and improves work efficiency. According to the traditional mode, the time for manual identification of indicators is about 110 / h, or other technologies require about 3 days of pre-work to identify indicators. If this technology is used, it only takes 1 day to train a large model.
[0041] High-accuracy matching extraction greatly reduces the matching error rate and greatly improves the accuracy rate. Due to the use of matching recognition, semantic understanding, interactive confirmation and other steps within a large model, the experimental test accuracy can reach 98.6%.
[0042] Industry personnel interactively confirm to ensure that the matching results meet expectations and reduce the risk of misunderstanding.
[0043] To reduce subjective bias and improve the objectivity and accuracy of decision-making, the experiment involves learning about 230 pieces of industry knowledge and decision-making documents. The test results are all objective answers, and the indicators will not be mixed with personal colors. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0045] Attached Figure 1It is a flowchart of semantic feature learning of a smaller model within a large model in the understanding matching extraction method based on the industry large model;
[0046] Attached Figure 2 It is a flowchart of the top items above the threshold in the understanding matching extraction method based on the industry big model;
[0047] Attached Figure 3 It is a flowchart of semantic understanding of context sentences in the understanding matching extraction method based on the industry big model;
[0048] Attached Figure 4 It is a schematic diagram of the optimal process of extracting semantic understanding results in the understanding matching extraction method based on the industry big model;
[0049] Attached Figure 5 It is a schematic diagram of the interactive confirmation process in the understanding matching extraction method based on the industry big model;
[0050] Attached Figure 6 It is a flowchart of exception handling in the understanding matching and extraction method based on the industry big model. DETAILED DESCRIPTION
[0051] In order to enable those skilled in the art to better understand the solution of the present invention, the present invention is further described in detail below in conjunction with specific implementation methods. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0052] A best embodiment is given below:
[0053] In this embodiment, a method for understanding, matching and extracting based on an industry big model has the following steps:
[0054] S1, the smaller models within the large model learn the semantic features of industry-specific indicators;
[0055] like Figure 1 As shown, further comprising:
[0056] S1-1. Determine the learning objectives, define the specific problems to be solved, the expected business goals, and the role of the big model in specific application scenarios;
[0057] For example, the goal may be to use a large model to classify customer purchase behavior for accurate sales, or to build a lightweight text sentiment analysis model for real-time monitoring of social media public opinion. Clarifying learning goals will help in the selection of special indicators, the construction of data sets, and the evaluation of large model performance in subsequent steps.
[0058] S1-2. Collect special indicators. Collect relevant special indicators according to learning objectives and large models.
[0059] These indicators should be able to fully reflect the performance of the large model on a specific task, such as precision, recall, F1 score in classification tasks, or mean square error, R in regression tasks. 2 Coefficients, etc.
[0060] Ensure that the selected indicators can reflect the core functions of the big model and meet business needs. For example, in the trade market industry, it may be necessary to collect special indicators such as industry personnel's browsing history, purchase frequency, product preferences, etc. to characterize the consumption characteristics of industry personnel.
[0061] S1-3. Determine whether the data is complete and conduct a comprehensive check on the collected data to confirm its completeness, including verifying whether the number of data records meets expectations, whether there are missing values for each special indicator, and whether the data distribution is reasonable.
[0062] For example, if it is found that the browsing history of people in a certain industry is missing a lot of key information, or the purchase data of a certain type of product is seriously insufficient, the data should be considered incomplete.
[0063] S1-4, data preprocessing (conditional branching), if the judgment result shows that the data is complete (that is, the result of the third step is "yes"), then enter the data preprocessing stage;
[0064] This stage includes data cleaning (removing outliers, filling missing values), normalization (such as normalization, standardization), feature engineering (deriving new features, dimensionality reduction, etc.), etc., aiming to improve the quality of data and make it more suitable as input for smaller models in the big model. For example, in the sales industry, standardizing the age of industry personnel so that they fall into the same scale range is conducive to the big model to fairly compare the consumption behavior of industry personnel of different age groups.
[0065] S1-5, use a smaller model within the larger model to learn semantic features. After completing data preprocessing, use the smaller model for training to learn the semantic features contained in the special indicators.
[0066] In the process of selecting and training large models, resource constraints such as memory usage and computational efficiency should be fully considered, while ensuring that the large model architecture matches the learning objectives and data characteristics. For example, a lightweight convolutional neural network (CNN) can be used to extract features from image data, or an embedded recurrent neural network (RNN) can be used to capture the dynamic semantics of time series data.
[0067] S1-6. Evaluate the effect of the large model. After the large model training is completed, use special indicators to conduct a detailed evaluation of the performance of the large model;
[0068] By comparing the indicator results on the training set, validation set, and test set, we analyze key factors such as the generalization ability, overfitting risk, and stability of the large model. For example, we observe the changes in the accuracy and recall rate of the large model on different data slices, and evaluate its recognition consistency for various industries and types of industry personnel groups.
[0069] S1-7, exception handling (loop logic). If the result of step S3 is that the data is incomplete (i.e., there is missing data), the exception handling process is triggered. At this time, the data source should be traced back to find the cause, or the missing data should be supplemented with the help of data interpolation technology. After that, steps S3 to S6 are re-executed to form an iterative loop until the data meets the integrity requirements and the large model learning and evaluation process can proceed smoothly.
[0070] To sum up, in the entire process from goal determination to model evaluation in the large-scale and small-scale model learning industry, it is necessary to emphasize data integrity judgment and its impact on subsequent steps, as well as the exception handling mechanism for data missing, to ensure that when encountering data problems, it can still proceed according to the predetermined path, and ultimately achieve effective learning of the semantic characteristics of special indicators in the industry and accurate evaluation of large model performance.
[0071] S2, set a similarity threshold and extract the top items whose semantic similarity is higher than the threshold;
[0072] like Figure 2 As shown in the figure, calculating the semantic similarity to judge the matching degree between texts is a prerequisite for obtaining high-quality results. First, set a similarity threshold, such as 0.75, as a basis for judgment. Then, use algorithms (such as cosine similarity, Jaccard similarity, etc.) to calculate the semantic similarity between the texts to be compared. Take two news reports as an example, and compare their overlap in topics, keywords, etc.
[0073] Then, a judgment is made. If the calculated similarity is higher than the preset threshold, it indicates that the semantics of the two texts are highly consistent. At this time, relevant indicators (such as the number of co-occurrences of keywords, quantitative indicators of topic consistency, etc.) can be further extracted for analysis. On the contrary, if the similarity is lower than the threshold, the two texts are judged to be semantically inconsistent. The large model feedbacks the conclusion of "semantic inconsistency" and throws out the abnormal reasons for the low similarity in detail, such as significant deviation of the topic, too large difference in core vocabulary, etc., to facilitate industry personnel to understand and adjust the query conditions or texts to be compared.
[0074] S3. The large model realizes the semantic understanding of the whole sentence, including context association and semantic inference;
[0075] like Figure 3As shown in the figure, the large model in semantic understanding achieves the purpose of deep semantic understanding of complete sentences input by industry practitioners through recursive analysis and context integration.
[0076] First, the big model receives a complete sentence provided by industry practitioners as initial input. Next, the big model makes preliminary semantic and contextual relevance judgments, and evaluates the relevance of the sentence to its known context (such as previous conversation history, domain knowledge base, etc.) by considering the sentence's grammatical parsing, word meaning analysis, and potential background knowledge.
[0077] For example, in the science education industry, if the big model can identify that the statement involves the scientific topic of biological adaptability and finds that it is closely related to the ecological content discussed previously, then the judgment result is relevant.
[0078] If the judgment result is relevant, the large model enters the semantic inference stage, using natural language understanding and reasoning techniques, such as dependency analysis and concept map construction, to generate an accurate understanding of the deep meaning of the sentence.
[0079] However, if the big model's judgment result shows that the sentence is irrelevant to the current context, the big model does not immediately give up understanding, but instead starts a context search mechanism to try to find context information related to the input sentence topic from a wider range of Internet industry resources (such as relevant literature, encyclopedia entries, network information, etc.).
[0080] Once the relevant context is found, the big model integrates it with the original context, expands the knowledge background, and then returns to step 2 to re-judge the relevance of the semantics with the updated context, forming an iterative cycle. The purpose of this is to gradually enrich the context and improve the ability to understand complex or novel sentences in a certain industry.
[0081] However, if after several such cycles (the specific number can be adjusted dynamically according to the big model settings or actual conditions), the big model still cannot find new contexts that are sufficient to support semantic understanding, or the newly introduced contexts still cannot establish an effective connection between the sentence and the existing knowledge system, then it is considered that effective semantic reasoning cannot continue.
[0082] At this point, the big model terminates the loop and feeds back an error message of "unable to perform semantic reasoning" to industry practitioners, and explains the difficulties encountered as much as possible, such as lack of professional knowledge support in a specific field, overly vague or ambiguous statements, etc., so that industry practitioners can adjust the way of asking questions or provide additional clues to promote subsequent effective communication.
[0083] S4. According to the semantic understanding result, the best matching item is selected as the extraction result;
[0084] like Figure 4As shown, the matching degree of the received semantic understanding results is evaluated, and the final extraction results or error messages are generated based on them, depending on the following process. First, the big model obtains the semantic understanding results generated by the previous links. For example, the parsing output of the query "global climate change trends in the past decade" by industry practitioners may include the keywords "global", "climate change", "trends in the past decade" and the logical relationships between them.
[0085] Next, the big model uses predefined matching algorithms or rules to calculate the matching degree between the understanding results and the internal knowledge base, database or specific target data set. For example, the key elements in the understanding results are compared with the time period, geographical range, climate indicators and other fields recorded in the climate database to obtain a quantitative matching value.
[0086] The key judgment point is that the big model compares the calculated matching degree with the preset threshold. If the matching degree reaches or exceeds the threshold, it means that the understanding result is highly consistent with the target data. The big model can select the best match, that is, the data record or information fragment that best matches the understanding result, to generate accurate extraction results. For example, in the weather forecasting industry, the big model may extract the change curves of the global average temperature, precipitation, etc. and a summary of significant climate events in the past decade as a detailed answer to the queries of industry personnel.
[0087] However, if the matching degree does not reach the threshold, it means that there is a large discrepancy between the semantic understanding result and the expected data source. It may be that the understanding result is vague or contains errors, or there is a lack of records in the target data that accurately correspond to the understanding result. In this case, the big model does not generate extraction results, but instead gives industry personnel feedback that "the semantic understanding result is vague or wrong", suggesting that industry practitioners rephrase the question, clarify key details, or inform the big model that it is currently unable to accurately respond to the query and needs to seek other information sources or manual assistance.
[0088] S5. Interactive confirmation to ensure that the matching results meet the intentions of industry personnel;
[0089] like Figure 5 As shown in the figure, after receiving the matching results, the big model interacts with industry practitioners to confirm their accuracy, and generates or adjusts the final result indicators based on the feedback of industry personnel. The process is as follows. First, the big model receives the preliminary matching results from the matching module. For example, in the product recommendation scenario, the big model matches a series of product lists that match the shopping history and preference characteristics of the industry personnel. Next, the big model presents these matching results to the industry personnel in an intuitive and easy-to-understand way, such as displaying product pictures, names, prices, and brief descriptions on the industry personnel interface. At the same time, the big model initiates a confirmation request, asking the industry personnel whether they think these matching results meet their needs, and the industry personnel give feedback by clicking the "yes" or "no" button.
[0090] The core loop judgment lies in the confirmation of the matching results by the industry personnel. If the industry personnel confirm that the matching results are accurate, that is, select "Yes", the big model will record the matching process and generate or update the final shopping indicators of the industry personnel based on the matching results, such as purchase conversion rate, industry personnel satisfaction score, etc. In addition, the big model may further record the specific selected product information for subsequent personalized recommendation optimization.
[0091] On the contrary, if the industry personnel feedback that the matching results are inaccurate (select "No"), the large model will start the loop processing mechanism, re-select other options from the alternative matching results that previously exceeded the preset threshold and show them to the industry personnel again.
[0092] This process will continue until a match is found that is approved by the industry personnel, or the industry personnel actively cancels the confirmation process. This loop design is designed to improve the experience of industry personnel, ensuring that industry personnel obtain satisfactory matching results, while also providing the large model with the opportunity to dynamically learn and adjust matching strategies.
[0093] S6. If the matching result does not meet the intention of the industry personnel, repeat the above steps to perform matching extraction;
[0094] like Figure 6 As shown in the figure, regarding exception handling, if the matching result that satisfies the industry personnel is not found during the cycle, or the industry personnel cancel the confirmation multiple times in a row, the big model can trigger an alarm mechanism to prompt the operation personnel to intervene and analyze the reasons. It may be that the industry personnel's demand description is unclear, the matching algorithm needs to be optimized, or there are insufficient related products in the product library. Through manual intervention and background data analysis, the big model can be continuously iterated and upgraded to improve matching accuracy and industry personnel satisfaction.
[0095] Based on the above method, an understanding matching extraction device based on an industry big model in this embodiment includes: at least one memory and at least one processor;
[0096] The at least one memory is used to store a machine-readable program;
[0097] The at least one processor is used to call the machine-readable program to execute a large model-based understanding matching extraction method.
[0098] The above-mentioned specific implementations are only specific cases of the present invention. The patent protection scope of the present invention includes but is not limited to the above-mentioned specific implementations. Any technical solutions that comply with the claims of the present invention and any appropriate changes or substitutions made by ordinary technicians in the relevant technical field shall fall within the patent protection scope of the present invention.
[0099] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for understanding, matching and extraction based on an industry big model, characterized in that: The steps are as follows: S1, the smaller models within the large model learn the semantic features of industry-specific indicators; S2, set a similarity threshold and extract the top items whose semantic similarity is higher than the threshold; S3. The large model realizes the semantic understanding of the whole sentence, including context association and semantic inference; S4. According to the semantic understanding result, the best matching item is selected as the extraction result; S5. Interactive confirmation to ensure that the matching results meet the intentions of industry personnel; S6. If the matching result does not meet the intention of the industry personnel, repeat the above steps to perform matching extraction.
2. According to the understanding matching extraction method based on the industry big model according to claim 1, it is characterized in that: In step S1, further comprising: S1-1. Determine the learning objectives, define the specific problems to be solved, the expected business goals, and the role of the big model in specific application scenarios; S1-2. Collect special indicators. Collect relevant special indicators according to learning objectives and large models. S1-3. Determine whether the data is complete and conduct a comprehensive check on the collected data, including verifying whether the number of data records meets expectations, whether there are missing values for each special indicator, and whether the data distribution is reasonable. S1-4, data preprocessing, if the judgment result shows that the data is complete, then enter the data preprocessing stage; S1-5, use a smaller model within the larger model to learn semantic features. After completing data preprocessing, use the smaller model for training to learn the semantic features contained in the special indicators. S1-6. Evaluate the effect of the large model. After the large model training is completed, use special indicators to conduct a detailed evaluation of the performance of the large model; S1-7, exception handling. If the result of step S3 is that the data is incomplete, the exception handling process is triggered. At this time, the data source should be traced back to find the cause, or the missing data should be supplemented with the help of data interpolation technology. After that, steps S3 to S6 are re-executed to form an iterative loop until the data meets the integrity requirements and the large model learning and evaluation process can proceed smoothly.
3. The understanding matching extraction method based on the industry big model according to claim 2 is characterized in that: In step S2, first, a similarity threshold is set as a basis for judgment, and then the semantic similarity between the texts to be compared is calculated using an algorithm; Then, a judgment is made. If the calculated similarity is higher than the preset threshold, it indicates that the semantics of the two texts are highly consistent. At this time, relevant indicators can be further extracted; On the contrary, if the similarity is lower than the threshold, the two texts are judged to be semantically inconsistent, and the large model feedbacks the conclusion of semantic inconsistency and throws out in detail the abnormal reasons that lead to low similarity.
4. The understanding matching extraction method based on the industry big model according to claim 3 is characterized in that: In step S3, first, the big model receives a complete sentence provided by an industry practitioner as the initial input. Next, the big model performs a preliminary semantic and contextual relevance judgment, and evaluates the relevance of the sentence to the known context by considering the sentence's grammatical parsing, word meaning analysis, and potential background knowledge.
5. The understanding matching extraction method based on industry big model according to claim 4 is characterized in that: If the judgment result is relevant, the big model will enter the semantic inference stage, using natural language understanding and reasoning technology. If the big model's judgment result shows that the sentence is irrelevant to the current context, the big model will not give up understanding immediately, but will start the context search mechanism to find context information related to the input sentence topic from a wider range of Internet industry resources; After finding the relevant context, the large model will be integrated with the original context to expand the knowledge background, and then return to step S2 to re-judge the relevance of the semantics with the updated context, forming an iterative loop.
6. The understanding matching extraction method based on the industry big model according to claim 5 is characterized in that: In step S4, first, the big model obtains the semantic understanding results generated by the previous link, and then the big model uses a predefined matching algorithm or rule to calculate the matching degree between the understanding results and the internal knowledge base, database or specific target data set; If the matching degree does not reach the threshold, it means that there is a large discrepancy between the semantic understanding result and the expected data source. At this time, the large model does not generate extraction results and provides feedback prompts to industry personnel.
7. The understanding matching extraction method based on the industry big model according to claim 6 is characterized in that: In step S5, first, the big model receives preliminary matching results from the matching module, and presents these matching results to the industry personnel. The big model initiates a confirmation request, asking the industry personnel whether they think these matching results meet the requirements, and the industry personnel give feedback by clicking a yes or no button; If the industry personnel confirms that the matching result is accurate, that is, they select Yes, the big model will record the matching process and generate or update the final shopping index of the industry personnel based on the matching result; If the industry personnel feedback that the matching result is inaccurate, that is, select "No", the large model will start the loop processing mechanism and re-select other options from the alternative matching results that previously exceeded the preset threshold and show them to the industry personnel again.
8. The understanding matching extraction method based on the industry big model according to claim 7 is characterized in that: In step S5, regarding exception handling, if a matching result that satisfies the industry personnel is not found during the loop process, or the industry personnel cancel the confirmation multiple times in a row, the large model triggers the alarm mechanism, prompting the operation personnel to intervene and analyze the cause.
9. An understanding matching extraction device based on an industry big model, characterized in that: include: at least one memory and at least one processor; The at least one memory is used to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute the method according to any one of claims 1 to 8.
Citation Information
Cited By
Self-supervision method and system for promoting chat robot to utilize unintegrated service
CN120849563A
